Reproduce “Meta-models for transfer learning in source localisation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Lawrence A. Bull, Matthew R. Jones, Elizabeth J. Cross, Andrew Duncan, Mark Girolami

Researchers from the University of Glasgow, University of Sheffield, Imperial College London, and University of Cambridge developed Bayesian multilevel meta-models to capture interdependencies between Acoustic Emission (AE) experiments. This framework enables prediction of model hyperparameters for new, unobserved systems, improving predictive performance in transfer learning and offering interpretable insights into how experimental design influences response characteristics.

Key insights

  • Analysis of independent single-task models revealed systematic and interpretable trends in AE response characteristics (GP process variance and length scale) with changes in experimental design parameters (e.g., sensor separation).
  • Multilevel models (MTL-B) can effectively capture these inter-task relationships, providing a more expressive representation of population-level behavior with quantified uncertainty compared to simpler shared prior models.
  • Modeling hyperparameters as functions of design variables offers deeper insights into experimental campaigns and provides a mechanism to encode weak domain expertise, which is more effective than directly modeling the raw response surface across tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.